# Cmux Browser Chatgpting

> This skill should be used when the user asks to "ask ChatGPT", "frag ChatGPT", "send a question to ChatGPT", "schick das an ChatGPT", "frag GPT ob", "ask GPT something", or wants to type a question into the ChatGPT browser surface open in cmux and read the response.

- Skill: `erich3000/cmux-browser-chatgpting` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add erich3000/cmux-browser-chatgpting`
- Raw SKILL.md: https://api.skillmd.com/api/skills/erich3000/cmux-browser-chatgpting/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Coding & Dev Tools
- Author: erich3000 (https://skillmd.com/u/erich3000)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/erich3000/cmux-browser-chatgpting

---


# cmux Browser ChatGPT

Send a question to the ChatGPT browser surface open in cmux and capture the response as a screenshot.

## Workflow

**Discovery:** Find the ChatGPT surface across all workspaces by scanning the cmux tree for `chatgpt.com`.

**Interaction:** Fill the prompt input, submit, wait for the response, screenshot.

Run the bundled script, passing the question as a single quoted argument:

```bash
bash <skill_base_dir>/scripts/ask-chatgpt.sh "<question>"
```

The script prints the surface ref it found and the path to the response screenshot.

After the script completes, read the screenshot and relay ChatGPT's answer to the user.

## Key Implementation Notes

- ChatGPT uses a `contenteditable` div with id `#prompt-textarea` — use `cmux browser fill`, not `type`
- `cmux browser eval` and `cmux browser press Enter` raise a JS exception on chatgpt.com — use the send-button click fallback chain instead
- `cmux browser snapshot` returns almost nothing on chatgpt.com (heavy JS app) — screenshots are the only reliable way to read the response
- Surface refs change between sessions — always discover fresh via `cmux tree --all`, never hardcode